How much do AI agents cost to run? Real numbers, explained

Running an AI agent costs whatever its model's tokens cost. For everyday jobs that's usually small: a morning inbox triage on a fast, small model costs a fraction of a cent, and a long research task on a top reasoning model costs a few cents to a few dimes. What drives the bill is how much the agent reads, how many steps it takes, and which model you pick.

Here's how to estimate it yourself.

Tokens, in one minute

Models read and write in tokens, roughly three-quarters of a word each. A 500-word email is about 650 tokens.

Every model has two prices, usually quoted per million tokens:

• Input: what the model reads (instructions, emails, tool results, the conversation so far).
• Output: what the model writes. Output usually costs several times more than input.

As of late 2026, small, fast models are commonly priced well under $1 per million input tokens, while frontier reasoning models are often several dollars per million input and more for output. Exact prices change often; Stormi's model picker shows each model's current price.

Why agents use more tokens than chat

An AI agent works in a loop. Each step, the model re-reads the instructions, the conversation, and every tool result so far, then decides what to do next. A task with eight tool calls means roughly eight model calls, each a little longer than the last.

Three things make agents expensive:

1. Big tool results. A search that returns 50 full emails is a lot of input.
2. Many steps. Every tool call is another round trip.
3. Screenshots. Images cost tokens too. (Stormi keeps only the last few in context.)

Worked examples

These are rough, to show the shape of the math. Your numbers depend on the model and your data.

Multiply by how often the job runs. A daily triage on a small model costs pennies a month. A daily deep digest on a frontier model can cost several dollars a month.

How to keep agent costs down

• Match the model to the job. Triage and summaries don't need the biggest model. Research and code review often do. In Stormi, each bot has its own model. See which AI model to use for agents.
• Give narrow jobs. "Triage unread email from the last 24 hours" reads far less than "look at my inbox."
• Schedule sensibly. A digest every Monday, not every hour.
• Watch usage. Stormi's Usage screen shows spend, requests, and tokens by day, bot, and model, so a bot that's burning money stands out.
• Set a limit. If you connect your own OpenRouter account, you can cap the key's spending there.

What Stormi costs

Stormi is free to download, with no subscription. Bots start on Stormi's own model account. Connect your OpenRouter account in Settings when you want to choose any model and pay providers directly. See pricing.